Merge branch 'branch-23.10' into processed_bytes_stream_distinct_coun…

…t_nvbench

conda/environments/all_cuda-118_arch-x86_64.yaml

@@ -9,6 +9,7 @@ channels:
 - nvidia
 dependencies:
 - aiobotocore>=2.2.0
+- aws-sdk-cpp<1.11
 - benchmark==1.8.0
 - boto3>=1.21.21
 - botocore>=1.24.21

conda/environments/all_cuda-120_arch-x86_64.yaml

@@ -9,6 +9,7 @@ channels:
 - nvidia
 dependencies:
 - aiobotocore>=2.2.0
+- aws-sdk-cpp<1.11
 - benchmark==1.8.0
 - boto3>=1.21.21
 - botocore>=1.24.21

conda/recipes/libcudf/conda_build_config.yaml

@@ -22,6 +22,9 @@ gbench_version:
 gtest_version:
   - ">=1.13.0"
 
+aws_sdk_cpp_version:
+  - "<1.11"
+
 libarrow_version:
   - "=12"
 

conda/recipes/libcudf/meta.yaml

@@ -74,6 +74,7 @@ requirements:
     - gtest {{ gtest_version }}
     - gmock {{ gtest_version }}
     - zlib {{ zlib_version }}
+    - aws-sdk-cpp {{ aws_sdk_cpp_version }}
 
 outputs:
   - name: libcudf
@@ -107,6 +108,7 @@ outputs:
         - dlpack {{ dlpack_version }}
         - gtest {{ gtest_version }}
         - gmock {{ gtest_version }}
+        - aws-sdk-cpp {{ aws_sdk_cpp_version }}
     test:
       commands:
         - test -f $PREFIX/lib/libcudf.so

cpp/benchmarks/io/cuio_common.cpp

@@ -15,6 +15,7 @@
  */
 
 #include <benchmarks/io/cuio_common.hpp>
+#include <cudf/detail/utilities/integer_utils.hpp>
 #include <cudf/detail/utilities/logger.hpp>
 
 #include <cstdio>
@@ -141,17 +142,18 @@ std::vector<std::string> select_column_names(std::vector<std::string> const& col
   return col_names_to_read;
 }
 
-std::vector<cudf::size_type> segments_in_chunk(int num_segments, int num_chunks, int chunk)
+std::vector<cudf::size_type> segments_in_chunk(int num_segments, int num_chunks, int chunk_idx)
 {
   CUDF_EXPECTS(num_segments >= num_chunks,
                "Number of chunks cannot be greater than the number of segments in the file");
-  auto start_segment = [num_segments, num_chunks](int chunk) {
-    return num_segments * chunk / num_chunks;
-  };
-  std::vector<cudf::size_type> selected_segments;
-  for (auto segment = start_segment(chunk); segment < start_segment(chunk + 1); ++segment) {
-    selected_segments.push_back(segment);
-  }
+  CUDF_EXPECTS(chunk_idx < num_chunks,
+               "Chunk index must be smaller than the number of chunks in the file");
+
+  auto const segments_in_chunk = cudf::util::div_rounding_up_unsafe(num_segments, num_chunks);
+  auto const begin_segment     = std::min(chunk_idx * segments_in_chunk, num_segments);
+  auto const end_segment       = std::min(begin_segment + segments_in_chunk, num_segments);
+  std::vector<cudf::size_type> selected_segments(end_segment - begin_segment);
+  std::iota(selected_segments.begin(), selected_segments.end(), begin_segment);
 
   return selected_segments;
 }

cpp/benchmarks/io/orc/orc_reader_options.cpp

@@ -19,6 +19,7 @@
 #include <benchmarks/io/cuio_common.hpp>
 #include <benchmarks/io/nvbench_helpers.hpp>
 
+#include <cudf/detail/utilities/integer_utils.hpp>
 #include <cudf/io/orc.hpp>
 #include <cudf/io/orc_metadata.hpp>
 #include <cudf/utilities/default_stream.hpp>
@@ -30,7 +31,7 @@
 constexpr int64_t data_size = 512 << 20;
 // The number of separate read calls to use when reading files in multiple chunks
 // Each call reads roughly equal amounts of data
-constexpr int32_t chunked_read_num_chunks = 8;
+constexpr int32_t chunked_read_num_chunks = 4;
 
 std::vector<std::string> get_top_level_col_names(cudf::io::source_info const& source)
 {
@@ -88,7 +89,7 @@ void BM_orc_read_varying_options(nvbench::state& state,
 
   auto const num_stripes =
     cudf::io::read_orc_metadata(source_sink.make_source_info()).num_stripes();
-  cudf::size_type const chunk_row_cnt = view.num_rows() / num_chunks;
+  auto const chunk_row_cnt = cudf::util::div_rounding_up_unsafe(view.num_rows(), num_chunks);
 
   auto mem_stats_logger = cudf::memory_stats_logger();
   state.set_cuda_stream(nvbench::make_cuda_stream_view(cudf::get_default_stream().value()));
@@ -99,7 +100,6 @@ void BM_orc_read_varying_options(nvbench::state& state,
       timer.start();
       cudf::size_type rows_read = 0;
       for (int32_t chunk = 0; chunk < num_chunks; ++chunk) {
-        auto const is_last_chunk = chunk == (num_chunks - 1);
         switch (RowSelection) {
           case row_selection::ALL: break;
           case row_selection::STRIPES:
@@ -108,7 +108,6 @@ void BM_orc_read_varying_options(nvbench::state& state,
           case row_selection::NROWS:
             read_options.set_skip_rows(chunk * chunk_row_cnt);
             read_options.set_num_rows(chunk_row_cnt);
-            if (is_last_chunk) read_options.set_num_rows(-1);
             break;
           default: CUDF_FAIL("Unsupported row selection method");
         }
@@ -132,9 +131,6 @@ using col_selections = nvbench::enum_type_list<column_selection::ALL,
                                                column_selection::ALTERNATE,
                                                column_selection::FIRST_HALF,
                                                column_selection::SECOND_HALF>;
-using row_selections =
-  nvbench::enum_type_list<row_selection::ALL, row_selection::STRIPES, row_selection::NROWS>;
-
 NVBENCH_BENCH_TYPES(BM_orc_read_varying_options,
                     NVBENCH_TYPE_AXES(col_selections,
                                       nvbench::enum_type_list<row_selection::ALL>,
@@ -146,6 +142,8 @@ NVBENCH_BENCH_TYPES(BM_orc_read_varying_options,
     {"column_selection", "row_selection", "uses_index", "uses_numpy_dtype", "timestamp_type"})
   .set_min_samples(4);
 
+using row_selections =
+  nvbench::enum_type_list<row_selection::ALL, row_selection::NROWS, row_selection::STRIPES>;
 NVBENCH_BENCH_TYPES(BM_orc_read_varying_options,
                     NVBENCH_TYPE_AXES(nvbench::enum_type_list<column_selection::ALL>,
                                       row_selections,

cpp/benchmarks/io/parquet/parquet_reader_options.cpp

@@ -19,28 +19,29 @@
 #include <benchmarks/io/cuio_common.hpp>
 #include <benchmarks/io/nvbench_helpers.hpp>
 
+#include <cudf/detail/utilities/integer_utils.hpp>
 #include <cudf/io/parquet.hpp>
 #include <cudf/utilities/default_stream.hpp>
 
 #include <nvbench/nvbench.cuh>
 
 // Size of the data in the benchmark dataframe; chosen to be low enough to allow benchmarks to
 // run on most GPUs, but large enough to allow highest throughput
-constexpr std::size_t data_size      = 512 << 20;
-constexpr std::size_t row_group_size = 128 << 20;
+constexpr std::size_t data_size = 512 << 20;
+// The number of separate read calls to use when reading files in multiple chunks
+// Each call reads roughly equal amounts of data
+constexpr int32_t chunked_read_num_chunks = 4;
 
 std::vector<std::string> get_top_level_col_names(cudf::io::source_info const& source)
 {
-  cudf::io::parquet_reader_options const read_options =
-    cudf::io::parquet_reader_options::builder(source);
-  auto const schema = cudf::io::read_parquet(read_options).metadata.schema_info;
-
-  std::vector<std::string> names;
-  names.reserve(schema.size());
-  std::transform(schema.cbegin(), schema.cend(), std::back_inserter(names), [](auto const& c) {
-    return c.name;
-  });
-  return names;
+  auto const top_lvl_cols = cudf::io::read_parquet_metadata(source).schema().root().children();
+  std::vector<std::string> col_names;
+  std::transform(top_lvl_cols.cbegin(),
+                 top_lvl_cols.cend(),
+                 std::back_inserter(col_names),
+                 [](auto const& col_meta) { return col_meta.name(); });
+
+  return col_names;
 }
 
 template <column_selection ColSelection,
@@ -55,6 +56,8 @@ void BM_parquet_read_options(nvbench::state& state,
                                                 nvbench::enum_type<UsesPandasMetadata>,
                                                 nvbench::enum_type<Timestamp>>)
 {
+  auto const num_chunks = RowSelection == row_selection::ALL ? 1 : chunked_read_num_chunks;
+
   auto constexpr str_to_categories = ConvertsStrings == converts_strings::YES;
   auto constexpr uses_pd_metadata  = UsesPandasMetadata == uses_pandas_metadata::YES;
 
@@ -87,9 +90,8 @@ void BM_parquet_read_options(nvbench::state& state,
       .use_pandas_metadata(uses_pd_metadata)
       .timestamp_type(ts_type);
 
-  // TODO: add read_parquet_metadata to properly calculate #row_groups
-  auto constexpr num_row_groups = data_size / row_group_size;
-  auto constexpr num_chunks     = 1;
+  auto const num_row_groups = read_parquet_metadata(source_sink.make_source_info()).num_rowgroups();
+  auto const chunk_row_cnt  = cudf::util::div_rounding_up_unsafe(view.num_rows(), num_chunks);
 
   auto mem_stats_logger = cudf::memory_stats_logger();
   state.set_cuda_stream(nvbench::make_cuda_stream_view(cudf::get_default_stream().value()));
@@ -100,18 +102,15 @@ void BM_parquet_read_options(nvbench::state& state,
       timer.start();
       cudf::size_type rows_read = 0;
       for (int32_t chunk = 0; chunk < num_chunks; ++chunk) {
-        auto const is_last_chunk = chunk == (num_chunks - 1);
         switch (RowSelection) {
           case row_selection::ALL: break;
           case row_selection::ROW_GROUPS: {
-            auto row_groups_to_read = segments_in_chunk(num_row_groups, num_chunks, chunk);
-            if (is_last_chunk) {
-              // Need to assume that an additional "overflow" row group is present
-              row_groups_to_read.push_back(num_row_groups);
-            }
-            read_options.set_row_groups({row_groups_to_read});
+            read_options.set_row_groups({segments_in_chunk(num_row_groups, num_chunks, chunk)});
           } break;
-          case row_selection::NROWS: [[fallthrough]];
+          case row_selection::NROWS:
+            read_options.set_skip_rows(chunk * chunk_row_cnt);
+            read_options.set_num_rows(chunk_row_cnt);
+            break;
           default: CUDF_FAIL("Unsupported row selection method");
         }
 
@@ -130,14 +129,26 @@ void BM_parquet_read_options(nvbench::state& state,
   state.add_buffer_size(source_sink.size(), "encoded_file_size", "encoded_file_size");
 }
 
+using row_selections =
+  nvbench::enum_type_list<row_selection::ALL, row_selection::NROWS, row_selection::ROW_GROUPS>;
+NVBENCH_BENCH_TYPES(BM_parquet_read_options,
+                    NVBENCH_TYPE_AXES(nvbench::enum_type_list<column_selection::ALL>,
+                                      row_selections,
+                                      nvbench::enum_type_list<converts_strings::YES>,
+                                      nvbench::enum_type_list<uses_pandas_metadata::YES>,
+                                      nvbench::enum_type_list<cudf::type_id::EMPTY>))
+  .set_name("parquet_read_row_selection")
+  .set_type_axes_names({"column_selection",
+                        "row_selection",
+                        "str_to_categories",
+                        "uses_pandas_metadata",
+                        "timestamp_type"})
+  .set_min_samples(4);
+
 using col_selections = nvbench::enum_type_list<column_selection::ALL,
                                                column_selection::ALTERNATE,
                                                column_selection::FIRST_HALF,
                                                column_selection::SECOND_HALF>;
-
-// TODO: row_selection::ROW_GROUPS disabled until we add an API to read metadata from a parquet file
-// and determine num row groups. https://github.com/rapidsai/cudf/pull/9963#issuecomment-1004832863
-
 NVBENCH_BENCH_TYPES(BM_parquet_read_options,
                     NVBENCH_TYPE_AXES(col_selections,
                                       nvbench::enum_type_list<row_selection::ALL>,

cpp/benchmarks/text/ngrams.cpp

@@ -36,11 +36,12 @@ static void BM_ngrams(benchmark::State& state, ngrams_type nt)
     cudf::type_id::STRING, distribution_id::NORMAL, 0, max_str_length);
   auto const column = create_random_column(cudf::type_id::STRING, row_count{n_rows}, profile);
   cudf::strings_column_view input(column->view());
+  auto const separator = cudf::string_scalar("_");
 
   for (auto _ : state) {
     cuda_event_timer raii(state, true);
     switch (nt) {
-      case ngrams_type::tokens: nvtext::generate_ngrams(input); break;
+      case ngrams_type::tokens: nvtext::generate_ngrams(input, 2, separator); break;
       case ngrams_type::characters: nvtext::generate_character_ngrams(input); break;
     }
   }

cpp/benchmarks/text/tokenize.cpp

@@ -67,8 +67,11 @@ static void bench_tokenize(nvbench::state& state)
       auto result = nvtext::count_tokens(input, cudf::strings_column_view(delimiters));
     });
   } else if (tokenize_type == "ngrams") {
-    state.exec(nvbench::exec_tag::sync,
-               [&](nvbench::launch& launch) { auto result = nvtext::ngrams_tokenize(input); });
+    auto const delimiter = cudf::string_scalar("");
+    auto const separator = cudf::string_scalar("_");
+    state.exec(nvbench::exec_tag::sync, [&](nvbench::launch& launch) {
+      auto result = nvtext::ngrams_tokenize(input, 2, delimiter, separator);
+    });
   } else if (tokenize_type == "characters") {
     state.exec(nvbench::exec_tag::sync,
                [&](nvbench::launch& launch) { auto result = nvtext::character_tokenize(input); });